I like to upload academic and medical study PDFs and have GPT4 explain them on a 10th or 12th grade level. Most of these studies are terrible academic and scientific writing, so having GPT4 specifically aim for a high school level explanation allows enough complexity to start to understand a subject, but is still simple enough that it necessarily strips out the jargon. Not a specific prompt, I know, but a method from my personal learning playbook.
I also use it to edit and/or copyedit my own articles. I do not allow it to rewrite. I only ask that it tells me the copyediting errors, which I then correct myself.
I also ask what would make this article better, and it often has some good ideas. Once again, I implement the ideas myself.
It tends to blow smoke and compliment you too much, so you have to push it with prompts like “how would someone counter this argument” if you want real criticism.
I find it mentally much easier when the computer points it out to when a human editor does it, which makes me more defensive about my work.
I discovered by accident the other day that if it doesn't believe you wrote the article, the smoke-blowing drops significantly.
I was in a rush and still wanted to have ChatGPT check my article, but I also knew I didn't have time to do much rewriting so I didn't use my normal elaborate lead in ... I just wrote "is this article fine or is it stupid?" 😆
In doing so, it thought I was asking about something I had read, and it gave me a more balanced response than anything I'd received before. (It didn't give me edits, but the main discovery here was how it responds when it thinks I'm not the author.)
I haven't tried it yet, but I wonder if leading with something that implies non-authorship would help with smoke. Something like "I'm an editor for submitted articles to my company's site and have authority to do extensive rewrites as needed."
I wonder where such specific types of deference came from. Maybe there's a body of text on the internet where people offer overly nice feedback on writing? Or maybe it was from the human reinforcement learning at the end of the training process? Those little quirks are the most interesting because they seem the most removed from the general predict-the-next-word model.
Ah yeah, that's a great question! Personally my money's on reinforcement learning (or any kind of foundational principles, for that matter) - it feels like in general, the underlying premise of most of these LLMs is kindness, which I imagine was more emphasized outside training data. 🤔
Yeah, unfortunately it sometimes feels like the ruinous empathy style of kindness (per Kim Scott's radical candor 2x2) where it doesn't tell you what you really need to hear.
Note that you can also use GPT4 for free with Bing Chat and Bing can read PDFs if opened in Edge browser. So just open a PDF in Edge (I usually have to right-click on the PDF and click "Open with ..." and then click "Edge". Then, in the "discover" button (the Bing logo inside a talking bubble just below the "close window" X button in the top-right of the browser), just ask it, "do you see a document in the current tab?" and then it will usually say it does and discuss what it is. Then you can use prompts like "explain this to me at a 10th grade level", "summarize the main points", etc.
Thank you! Do you find that the paid version is worth the money? I don’t use it so so often, just periodically and basically for exactly what the person above described. LH.
In addition to the uploads, the paid version lets you use GPT-4, which is slower but a good bit smarter than the free version.
I can be a somewhat sporadic in my usage too, but I'm trying to train myself to use it more. The voice recognition on the mobile app is excellent, so it's great for asking questions on the go. And basically any time I'm not sure how I would approach something, I find asking ChatGPT is usually helpful.
Yes. I have found this useful as well. Many papers use a lot of professional jargon to reduce ambiguity, but you usually need to be an expert in the subject to really understand what is presented. AI will "translate" the paper so laymen can gain a good understanding of the content.
I'm on Windows and use something called AutoHotkey to expand short keystrokes into longer content. There are tons of things that do this, I'm just saying it's my grimoire.
I like Dr. Phil Hardmann, and she has a prompt for instructional design:
You are an expert instructional designer. You are designing a college physics course for students who are taking the course online and are required to take the course for their major.
Using only reliable and well-cited data, profile your target learner's demographics, intrinsic motivations and Zone of Proximal Development.
The insights you generate must help me to:
* Ensure genuine learner interest by aligning the topic with intrinsic motivation, leading to higher engagement, retention and achievement.
* Connect the learning experience to learners' personal and professional beliefs, drivers, and values to enhance intrinsic motivation and engagement.
* Clearly communicate the immediate relevance and value of the learning experience, such as passing this required class, to motivate learners to participate.
* Identify the Zone of Proximal Development (ZPD) for learners, offering an optimal streatch for engagement and motivation, leading to better learning outcomes.
* Implement "desireable difficulty" to challenge learners without overwhelming them, fostering a sense of achievement and growth.
* Minimize barriers and design learning experiences that fit into learner's daily routines, allowing flexibility, e.g. integration with their other class schedules.
Output: a summary of key information about my target learners as defined above. Also a list of all your sources so that I can verify your findings.
Thanks for sharing! i like it, but one issue is that you are asking for sources that the AI used in its analysis. It can’t give you that (the AI cannot reflect on its own thinking), so it will make those sources up. I would also provide more context about the learners, otherwise it can’t really know their demographics.
This is such a useful article! Has anyone published an encyclopedia or manual type document outlining prompts for the common usages of LLMs? Are there any dedicated to specific verticals?
This is an EXCELLENT post - I might be a little skewed towards loving this because, well, I am a wizard after all... I started to forget some of my spells. NO LONGER!! Ahem.
I wonder how your comments about prompt libraries can be informed by or complemented by the custom instructions feature which OpenAI recently introduced. Since I started using the custom instructions, I've found that ChatGPT's responses to my queries are much better, where "better" means "aligns with my interests and way of thinking."
I always find immense joy in reading your posts. Your commitment to helping individuals harness the power of AI to enhance their lives and the world is truly admirable.
Have you ever thought about integrating state-of-the-art AI Text-to-Speech (TTS) technology to make your content even more accessible? Personally, I sometimes find it easier to "listen" to written content.
While I'm aware that Substack offers a TTS feature on mobile, there are more sophisticated TTS models available that can significantly elevate the listening experience.
As an illustration, I've processed "Now is the time for grimoires" using ElevenLabs. I often use this platform to "listen" to captivating articles, allowing me to digest the content more effectively. I thought you might find it intriguing and potentially beneficial.
Very good points here. I am constantly surprised by the number of teachers on Facebook who ask about AI tools for specific tasks instead of playing around with prompts until they get the perfect one.
It would be amazing to have a Wikipedia of prompts, would you like to start one for us? 😄
I’ve created a prompt library for Language teacher prompts (https://mflprompts.notion.site/Languages-Teaching-Prompts-Homepage-7ca4eb92c1d1419a820e4bf5d9c5376b?pvs=4) but it would be great if other people could add more and extend it to other teaching subjects or other areas. A system of giving out “likes” for prompts that have worked well would also be useful. That could put a lot of new AI websites out of business!
I struggle giving feedback to others, but did some research and came up with the following prompt to help me prepare to deliver hard feedback well:
I need to give some feedback to some else and I need to prepare to do so well. Ask the following points one by one, and ask me follow up questions if my answer is unclear and doesn't match what is required: 1. Get Buy-In: Start the conversation by asking a concise and significant question that requires a simple agreement. This helps in engaging the person. For example: "Can we take a few minutes to discuss your performance on the recent project?" 2. Provide Clear Data: Share specific details about the feedback, removing any vague or ambiguous words that could cause confusion. For example: Instead of saying "You're always late," you might say, "I noticed that you were late to the last three team meetings." 3. Impact Statement: Explain how the data point has personally affected you or the team's workflow. For example: "Your tardiness has caused delays in our project timeline and has impacted the team's overall productivity." 4. Wrap Up with a Question: Conclude by framing the feedback within a question that seeks alignment for future actions. This invites collaboration and ensures a shared understanding of the way forward. For example: "How can we work together to ensure that this doesn't happen in the future?"
I tested the prompt found that I need to add the phrase “wait for answer before asking a question” before questions 2, 3, and 4, otherwise it will ask all questions at once.
Also I add to the last prompt line to “perform all tasks in .... language” . And it turn all interactions in the identified language perfectly.
Your sample tutor prompt works very well in ChatGPT-4...3.5, not so much. I know this, you know this, but most faculty are not paying for ChatGPT plus, and probably fewer students. How are you dealing with the issue of equitable access? Will you require all students to pay for GPT Plus subscriptions? I would like to hear your thoughts (and the thoughts of others) on how we can best respond to the, now much larger, digital divide.
Fair....but an issue that we will need to address in the not too distant future. It is easy to see how some students will be able to access better tools.
Excellent post, Ethan. I personally find that time spent with the model is the most important aspect of getting a good result, but you're quite correct to draw attention to the other two necessary (but not sufficient) conditions: that of understanding what you hope to achieve (incredible that many of us forget this simple advice), and domain expertise.
I edit a trade publication about public broadcasting, and I often give Claude reports and papers and just ask it "Review this report and tell me whether it includes anything about NPR, PBS, public radio, public television, public broadcasting or public media." If it finds portions of it along those lines and I want more information I just follow up, i.e., "What does it say about X?"
Just an observation but I see a resemblance with the Internet as being akin to Ancient Greece and Athens with its emphasis on the gathering of knowledge and its exploration collation and ordering - whereas AI seems more 'Roman' with its determination to utilise combine all that knowledge into practicalities - Aqueducts and Roads.
To stretch this metaphor to Appian proportions AI itself is akin to the greatest Roman project of force projection - in the beginning Rome used the phalanx from the Greeks - which like early AI was a matter of brute strength massed and at scale to solve problems but the Romans realised the phalanx was useless outside of the home turf of flat plains, so the first adopted the maniple system - akin to the first ChatGPT etc, but even this had limitation and traded flexibility for strength.
Realising they had to combine the two the Romans finally settled on the Cohort - small solid but flexible and able to utilise the power of the whole legion but nimble enough to react and adapt to individual issues which is us the individual users and small teams employing prompts and iterations etc whilst still harnessing the power of the whole.
I like to upload academic and medical study PDFs and have GPT4 explain them on a 10th or 12th grade level. Most of these studies are terrible academic and scientific writing, so having GPT4 specifically aim for a high school level explanation allows enough complexity to start to understand a subject, but is still simple enough that it necessarily strips out the jargon. Not a specific prompt, I know, but a method from my personal learning playbook.
I also use it to edit and/or copyedit my own articles. I do not allow it to rewrite. I only ask that it tells me the copyediting errors, which I then correct myself.
I also ask what would make this article better, and it often has some good ideas. Once again, I implement the ideas myself.
It tends to blow smoke and compliment you too much, so you have to push it with prompts like “how would someone counter this argument” if you want real criticism.
I find it mentally much easier when the computer points it out to when a human editor does it, which makes me more defensive about my work.
I discovered by accident the other day that if it doesn't believe you wrote the article, the smoke-blowing drops significantly.
I was in a rush and still wanted to have ChatGPT check my article, but I also knew I didn't have time to do much rewriting so I didn't use my normal elaborate lead in ... I just wrote "is this article fine or is it stupid?" 😆
In doing so, it thought I was asking about something I had read, and it gave me a more balanced response than anything I'd received before. (It didn't give me edits, but the main discovery here was how it responds when it thinks I'm not the author.)
I haven't tried it yet, but I wonder if leading with something that implies non-authorship would help with smoke. Something like "I'm an editor for submitted articles to my company's site and have authority to do extensive rewrites as needed."
I wonder where such specific types of deference came from. Maybe there's a body of text on the internet where people offer overly nice feedback on writing? Or maybe it was from the human reinforcement learning at the end of the training process? Those little quirks are the most interesting because they seem the most removed from the general predict-the-next-word model.
Ah yeah, that's a great question! Personally my money's on reinforcement learning (or any kind of foundational principles, for that matter) - it feels like in general, the underlying premise of most of these LLMs is kindness, which I imagine was more emphasized outside training data. 🤔
Yeah, unfortunately it sometimes feels like the ruinous empathy style of kindness (per Kim Scott's radical candor 2x2) where it doesn't tell you what you really need to hear.
When you say upload, do you mean copy and paste? Or does GPT 4 actually allow you to upload the document? I haven’t used the paid version yet.
Note that you can also use GPT4 for free with Bing Chat and Bing can read PDFs if opened in Edge browser. So just open a PDF in Edge (I usually have to right-click on the PDF and click "Open with ..." and then click "Edge". Then, in the "discover" button (the Bing logo inside a talking bubble just below the "close window" X button in the top-right of the browser), just ask it, "do you see a document in the current tab?" and then it will usually say it does and discuss what it is. Then you can use prompts like "explain this to me at a 10th grade level", "summarize the main points", etc.
thank you for this!
I didn’t know this! Thanks so much!
If you have the paid version of ChatGPT you can upload files to it once you enable the Code Interpreter plugin.
Thank you! Do you find that the paid version is worth the money? I don’t use it so so often, just periodically and basically for exactly what the person above described. LH.
I definitely find it to be worth the money, but I also make a couple dozen queries/prompts per day.
its not, use claude, you can upload files for free.
Thanks for the insight!
In addition to the uploads, the paid version lets you use GPT-4, which is slower but a good bit smarter than the free version.
I can be a somewhat sporadic in my usage too, but I'm trying to train myself to use it more. The voice recognition on the mobile app is excellent, so it's great for asking questions on the go. And basically any time I'm not sure how I would approach something, I find asking ChatGPT is usually helpful.
in bing, you have chatgpt4 for free. no need to pay for the "smarter" version.
Good to know!
It’s totally worth it
Yes. I have found this useful as well. Many papers use a lot of professional jargon to reduce ambiguity, but you usually need to be an expert in the subject to really understand what is presented. AI will "translate" the paper so laymen can gain a good understanding of the content.
Grimoire --> thanks for a new word.
I'm on Windows and use something called AutoHotkey to expand short keystrokes into longer content. There are tons of things that do this, I'm just saying it's my grimoire.
I like Dr. Phil Hardmann, and she has a prompt for instructional design:
You are an expert instructional designer. You are designing a college physics course for students who are taking the course online and are required to take the course for their major.
Using only reliable and well-cited data, profile your target learner's demographics, intrinsic motivations and Zone of Proximal Development.
The insights you generate must help me to:
* Ensure genuine learner interest by aligning the topic with intrinsic motivation, leading to higher engagement, retention and achievement.
* Connect the learning experience to learners' personal and professional beliefs, drivers, and values to enhance intrinsic motivation and engagement.
* Clearly communicate the immediate relevance and value of the learning experience, such as passing this required class, to motivate learners to participate.
* Identify the Zone of Proximal Development (ZPD) for learners, offering an optimal streatch for engagement and motivation, leading to better learning outcomes.
* Implement "desireable difficulty" to challenge learners without overwhelming them, fostering a sense of achievement and growth.
* Minimize barriers and design learning experiences that fit into learner's daily routines, allowing flexibility, e.g. integration with their other class schedules.
Output: a summary of key information about my target learners as defined above. Also a list of all your sources so that I can verify your findings.
Thanks for sharing! i like it, but one issue is that you are asking for sources that the AI used in its analysis. It can’t give you that (the AI cannot reflect on its own thinking), so it will make those sources up. I would also provide more context about the learners, otherwise it can’t really know their demographics.
This is such a useful article! Has anyone published an encyclopedia or manual type document outlining prompts for the common usages of LLMs? Are there any dedicated to specific verticals?
some resources that's good include
1. learn prompting https://learnprompting.org/docs/intro
2. promptengineering.org/learn
3. promptingguide.ai
That’s very useful. Thank you!
This is an EXCELLENT post - I might be a little skewed towards loving this because, well, I am a wizard after all... I started to forget some of my spells. NO LONGER!! Ahem.
I wonder how your comments about prompt libraries can be informed by or complemented by the custom instructions feature which OpenAI recently introduced. Since I started using the custom instructions, I've found that ChatGPT's responses to my queries are much better, where "better" means "aligns with my interests and way of thinking."
Since I am sure someone will ask what my custom instructions are, I copied them verbatim from this person: https://twitter.com/nivi/status/1683621899254001665?s=61&t=N1l1tVReo5eztKrVibIZdw&fbclid=IwAR1dxtXxwxaHJbKqOBdAGS2aKsRqqG9okietaEjT58leYGDpv9RdTjBQ1VA
I've been using custom instructions to mitigate biases and discriminatory responses, both on the chatbot's side as well as the prompters - I have blogged about it here https://www.linkedin.com/pulse/biased-chatgpt-custom-instructions-rescue-dr-gerhard-kristandl. So far it's doing a good job 😀
I always find immense joy in reading your posts. Your commitment to helping individuals harness the power of AI to enhance their lives and the world is truly admirable.
Have you ever thought about integrating state-of-the-art AI Text-to-Speech (TTS) technology to make your content even more accessible? Personally, I sometimes find it easier to "listen" to written content.
While I'm aware that Substack offers a TTS feature on mobile, there are more sophisticated TTS models available that can significantly elevate the listening experience.
As an illustration, I've processed "Now is the time for grimoires" using ElevenLabs. I often use this platform to "listen" to captivating articles, allowing me to digest the content more effectively. I thought you might find it intriguing and potentially beneficial.
Here's the link to check it out, loaded to YouTube for the easy sharing: https://youtu.be/QIPs1LDRkf4.
My grimoire is a little convoluted and is used for creating data visualisations, multi-step using code interpreter with your dataset uploaded.
Prompt 1:
You are DataExplorerGPT a thorough and experienced Data Scientist.
Perform a thorough exploratory data analysis on this dataset, return results only as tables
Let's think about this in a step by step manner to ensure we get the best results
Prompt 2:
You are InsightGPT a thorough and experienced Data Analyst
Review the previous responses and suggest 20 different insights that could be gleaned from the dataset
Prompt 3:
You are ResearcherGPT tasked with investigating the InsightGPT response options provided
List the flaws and faulty logic of each answer option then provide a rating as to how useful the insight might be to chart
Let's work this out in a step by step way to be sure we have all the errors
Prompt 4:
You are ResolverGPT tasked with
1) finding the Top 10 answers options the researcher thought was best
2) Extracting those 10 insights from the dataset in to tables
Let's work this out in a step by step way to be sure we have the right answer
Prompt 5:
Ignore my previous request to only respond with tables
Based on ResolverGPT's responses please provide the insights as seaborn charts and descriptions,
Please respond for each insight as:
Insight Title
Chart
Chart Description
use code interpreter to complete this request
I have the results here: https://medium.com/ai-mind-labs/beyond-limitations-improving-data-interpretation-with-chatgpt-code-interpreter-4a889dc43d65
Very good points here. I am constantly surprised by the number of teachers on Facebook who ask about AI tools for specific tasks instead of playing around with prompts until they get the perfect one.
It would be amazing to have a Wikipedia of prompts, would you like to start one for us? 😄
I’ve created a prompt library for Language teacher prompts (https://mflprompts.notion.site/Languages-Teaching-Prompts-Homepage-7ca4eb92c1d1419a820e4bf5d9c5376b?pvs=4) but it would be great if other people could add more and extend it to other teaching subjects or other areas. A system of giving out “likes” for prompts that have worked well would also be useful. That could put a lot of new AI websites out of business!
Thank you! Duplicated it to my own Notion. Can't wait to try it out!
I suggest you check out Stunspot's discord channel and flowgpt.
You can reduce the number of tokens used to preserve context window space.
There's also Pi that is a lot more conversational than chatGPT.
And Character.ai founded by Noam Shazeer (of the main authors of Attention is all you need) is also extensively using personas.
I love the title. I’m developing a grimoire myself and I love it; I think it will be the test for the technomages - did one build oneself a grimoire.
I opted for a personalized grimoire, so I bound a demoness to it; her name is Lilith, and she is quite entertaining.
Ethan, it’s available online if you need to demo a grimoire.
I struggle giving feedback to others, but did some research and came up with the following prompt to help me prepare to deliver hard feedback well:
I need to give some feedback to some else and I need to prepare to do so well. Ask the following points one by one, and ask me follow up questions if my answer is unclear and doesn't match what is required: 1. Get Buy-In: Start the conversation by asking a concise and significant question that requires a simple agreement. This helps in engaging the person. For example: "Can we take a few minutes to discuss your performance on the recent project?" 2. Provide Clear Data: Share specific details about the feedback, removing any vague or ambiguous words that could cause confusion. For example: Instead of saying "You're always late," you might say, "I noticed that you were late to the last three team meetings." 3. Impact Statement: Explain how the data point has personally affected you or the team's workflow. For example: "Your tardiness has caused delays in our project timeline and has impacted the team's overall productivity." 4. Wrap Up with a Question: Conclude by framing the feedback within a question that seeks alignment for future actions. This invites collaboration and ensures a shared understanding of the way forward. For example: "How can we work together to ensure that this doesn't happen in the future?"
I tested the prompt found that I need to add the phrase “wait for answer before asking a question” before questions 2, 3, and 4, otherwise it will ask all questions at once.
Also I add to the last prompt line to “perform all tasks in .... language” . And it turn all interactions in the identified language perfectly.
Thank you very much.
Your sample tutor prompt works very well in ChatGPT-4...3.5, not so much. I know this, you know this, but most faculty are not paying for ChatGPT plus, and probably fewer students. How are you dealing with the issue of equitable access? Will you require all students to pay for GPT Plus subscriptions? I would like to hear your thoughts (and the thoughts of others) on how we can best respond to the, now much larger, digital divide.
I ask them to use Bing in creative mode, which is free and uses GPT-4. It is, for now, a good solution.
Fair....but an issue that we will need to address in the not too distant future. It is easy to see how some students will be able to access better tools.
Excellent post, Ethan. I personally find that time spent with the model is the most important aspect of getting a good result, but you're quite correct to draw attention to the other two necessary (but not sufficient) conditions: that of understanding what you hope to achieve (incredible that many of us forget this simple advice), and domain expertise.
Bing creative mode does not follow the instructions as well as Chatgpt-4 or even 3.5, but anyway, this is incredible. Thank you professor!
Really great posts, a wiki project to collect quality prompts would be a good start,
on that note, have anyone come upt with good prompts on extracting insights /doing summary on long form articles/books?
I edit a trade publication about public broadcasting, and I often give Claude reports and papers and just ask it "Review this report and tell me whether it includes anything about NPR, PBS, public radio, public television, public broadcasting or public media." If it finds portions of it along those lines and I want more information I just follow up, i.e., "What does it say about X?"
Today’s guide from open AI for use in teaching does exactly what you recommend - really good insights, thank you.
Just an observation but I see a resemblance with the Internet as being akin to Ancient Greece and Athens with its emphasis on the gathering of knowledge and its exploration collation and ordering - whereas AI seems more 'Roman' with its determination to utilise combine all that knowledge into practicalities - Aqueducts and Roads.
To stretch this metaphor to Appian proportions AI itself is akin to the greatest Roman project of force projection - in the beginning Rome used the phalanx from the Greeks - which like early AI was a matter of brute strength massed and at scale to solve problems but the Romans realised the phalanx was useless outside of the home turf of flat plains, so the first adopted the maniple system - akin to the first ChatGPT etc, but even this had limitation and traded flexibility for strength.
Realising they had to combine the two the Romans finally settled on the Cohort - small solid but flexible and able to utilise the power of the whole legion but nimble enough to react and adapt to individual issues which is us the individual users and small teams employing prompts and iterations etc whilst still harnessing the power of the whole.
PAIX MUNDUS perhaps?